应用我们提出的构建块,用SPD-Conv替换了四个跨步卷积;但另一方面,我们只是删除了最大池化层,因为我们的主要目标是低分辨率图像,我们实验中使用的数据集的图像相当小(Tiny ImageNet 中为64 × 64 64 × 6464×64,CIFAR-10中为32 × 32 32 × 3232×32)因此不需要池化, 对于更大的图像,这样的最大池化层仍然...
This repository is no longer maintained. It has been relocated toSPD-Conv, where active updates and maintenance will continue. This repo contains the source code and evaluation scripts for our ECML PKDD 2022 paper. No More Strided Convolutions or Pooling: A New CNN Building Block for Low-Reso...